# Qwen3 Embedding 8B

> 4096-dimension embeddings with a 32K input window and strong multilingual retrieval. Use when recall quality matters more than the per-million rate, or when documents are long enough that a 2K window would have to chunk them.

Human version: https://kymaapi.com/models/qwen3-embedding-8b
Live JSON: `GET https://kymaapi.com/v1/models` (no auth required)

## Facts

- **Model ID**: `qwen3-embedding-8b` — pass this as `model` in the request body
- **Creator**: Alibaba
- **Released**: 2025-06-05
- **Context window**: 33K tokens
- **Price**: $0.0135 in / $0.00 out per 1M, cached input 10%
- **Capabilities**: tools, caching

## Call it

```bash
curl https://kymaapi.com/v1/chat/completions \
  -H "Authorization: Bearer $KYMA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "qwen3-embedding-8b", "messages": [{"role": "user", "content": "Hello"}]}'
```

## Positioning

Qwen3 Embedding 8B generates 4096-dimensional text embeddings optimized for multilingual retrieval and long-context RAG pipelines. Reach for it when document length or cross-language recall outweighs the need for the lowest possible per-token rate.

## About

This model produces dense vector representations with a 32,768-token context window. It is designed for text-only embedding tasks and does not support reasoning, vision, or structured output generation.

On Kyma, it runs as an OpenAI-compatible endpoint behind a single API key. Requests benefit from automatic failover if a serving path degrades, and responses return the exact cost in usage.cost alongside an X-Kyma-Model header. Prompt caching is supported, billing repeated prefixes at 10% of the standard input rate.

Because it is strictly an embedding model, it returns zero output tokens and cannot generate conversational text. It operates at a medium speed tier, making it better suited for batch indexing or retrieval-heavy workflows than for real-time, low-latency interactive search.

## Use cases

- **Multilingual Document Search** — Finds relevant passages across different languages without translation overhead.
- **Long-Form Context Indexing** — Embeds full documents up to 32K tokens to avoid aggressive chunking.
- **High-Recall Retrieval Pipelines** — Prioritizes semantic match quality over minimal compute cost for RAG systems.

## Not ideal for

Do not use this model for real-time chat, text generation, or tasks requiring sub-100ms latency, as it only outputs vectors and runs at a medium speed tier.

## Pick something else when

- You need minimal compute cost and fast indexing → `embeddinggemma-300m`

## See also

- All models: https://kymaapi.com/models.md
- Pricing: https://kymaapi.com/pricing.md
- Other models by Alibaba: https://kymaapi.com/models?q=Alibaba